Translation Costs Pennies, Rights Cost a Fortune
The economics of translating books with LLMs, translation rights, author payouts, and whether it’s possible to build a streaming service for books with automatic translation into multiple languages.
The Idea and the Main Problem
If you use Amazon Kindle, you probably know about Kindle Unlimited. You pay for a subscription and can read any books included in it. I had an idea to build something similar as a SaaS, but with the main feature being book translation into many different languages. The translation itself could be done with LLMs.
At first it sounds simple. You take a book, run it through a model, get the translation and let the user read it in the language they want. But the main problem is not even the translation. It is the rights to the book. A single book can have many different rights that the owner can sell separately, such as publishing rights, translation rights, adaptation rights, audiobook rights and so on. Translation rights can also be sold separately for each language. For example, Ukrainian translation rights might belong to one publisher, Polish rights to another one, and German rights to a third one.
And this is probably the biggest problem for a service like this. Especially because rights to popular books do not always stay with the author. They are often managed by publishers or literary agents that already handle licensing, promotion and distribution.
Paying Authors
The second problem is how to pay authors fairly. Paying simply per page does not make much sense because an ebook does not really have a fixed page count. You can increase the font size, make the content area smaller, change spacing, and suddenly the same book can have 700 pages instead of 300.
Amazon solves this with its own metric called KENPC, or Kindle Edition Normalized Page Count. Basically, Amazon normalizes the formatting of the book, including font size, margins, line height and other parameters, and gets a standardized page count. In the subscription, the author gets paid based on how many pages users actually read. The same page from the same user is counted only once. If someone reads the same page ten times, the author does not get paid ten times for it.
A similar model could work for this kind of SaaS. For example, a user pays $10 to $15 per month. All subscription revenue goes into one pool. The platform takes its commission, and the rest is distributed between rights holders based on how much normalized content from their books users actually read. It does not even have to be based on pages. You could count normalized characters, words or tokens instead.
Translation with LLMs
Building a workflow for book translation with LLMs is already much cheaper than paying a professional translator or a translation studio. Top models can already translate normal text quite well even without fine-tuning, and the quality is improving very fast. At this point it feels more like a matter of months than years.
Of course, simply taking a book and telling a model to translate it in one prompt is not enough. A proper workflow should at least build a glossary of terms and names so the translation stays consistent across the whole book. For example, memory map should not be translated as one thing in one chapter and something different in another. And bus in a technical book should not suddenly become a literal bus. The same applies to character names, place names, fictional terms, forms of address and dialogue style.
The translation can also be done in multiple passes. The first pass could be the rough translation. The second could check terminology and context. The third could focus on style and find weird or unnatural parts. Even with three passes, it can still be much cheaper than a full human translation.
If we take a model like GPT-6 Astra with API pricing of $10 per 1 million input tokens and $50 per 1 million output tokens, translating a 100,000 word book could cost around $10 for one full pass. Three passes would be around $30. For comparison, professional translation can easily cost several cents per word and sometimes around $0.10 per word. That means a 100,000 word book can cost thousands of dollars to translate, and in some cases close to $10,000.
The difference is huge. Of course, LLM translation is still not equal to a good literary translator, especially when it comes to fiction, humor, wordplay, cultural references or a very specific writing style. But the economics already look very interesting. I think book translation and translation in general will mostly move toward LLMs in the future.
Budget
Even if one translation costs something like $10 to $30, it scales very quickly. Let’s say the service wants to launch with at least 100 books, and each book needs to be translated into five popular languages. Even with a very optimistic estimate of $10 per language, that is already 100 books × 5 languages × $10 = $5,000.
And if you use multiple passes, add editing, quality checks and regenerate problematic parts, the budget can become much larger. And that is only translation. You still have servers, reader development, mobile apps, payments, licensing and legal work. So even if AI makes translation much cheaper, it does not mean you can launch this kind of service with $500.
Conclusion
LLMs are getting closer to translators than ever before. Human translation is expensive and slow, but still usually better in terms of quality. LLM translation is cheap and fast, but it can still lose on style, nuance and consistency. But looking at how quickly models are improving, I have very little doubt that most large scale text translation in the future will be done with LLMs.
The idea itself seems interesting to me, but AI is probably the smallest problem here. Everyone will have access to good models, so translation itself is unlikely to become a serious advantage. The much harder part is building a good catalog of books and getting the rights to translate and distribute them. Especially if you are a small service that nobody knows. On top of that, translation rights for popular books and major languages may already have been sold to other publishers, which means you simply cannot add those books to your catalog even if technically you can translate them in a few minutes.
So technically, building a product like this is already relatively realistic today. Legally and commercially, it is much more difficult. And that, not the quality of LLM translation, is probably the real problem with this kind of service.